a16z · Published 2026-06-08

Foundation Models are a Commodity | Benedict Evans on a16z

privateReview: unreviewedConfidence: high
foundation modelscommoditizationcapexinfrastructureproduct-market fitfoundation-modelsvalue-captureagentic-codingenterprise-aiexception-handlingmodel-abstraction
Open on YouTube ↗

Summary

Overview

  • Speaker: Benedict Evans and Erik Torenberg
  • Channel: a16z
  • Main topic: The economic, infrastructure, and strategic dynamics of AI foundation models and LLMs, drawing parallels to historical platform shifts like mobile, internet, and PCs.
  • Purpose: Analyze the state of generative AI, foundation model commoditization, infrastructure capex, and what historical platform shifts teach us about the current AI boom. Technology analyst Benedict Evans joins a16z General Partner Erik Torenberg to discuss his presentation 'AI Eats the World.' They examine why foundation models risk becoming commodities, the severe supply/demand imbalances and capex constraints in AI infrastructure, and the historical analogies between current generative AI adoption and past platform shifts like mobile and internet data. Evans argues that while software development has found a strong product-market fit with coding assistants, predicting the broader enterprise and consumer application landscape remains deeply uncertain.

Topic Map

The State of AI Product-Market Fit and Coding

  • Explanation: Evaluating where AI has achieved true product-market fit, focusing heavily on software coding assistants.
  • Key claims:
    • Agentic coding started working immediately, narrowing tech focus onto that specific use case.
    • Software development is the prime example of strong product-market fit because customers actively pull the tool out of your hands.
    • Most other enterprise and consumer use cases remain exploratory with open-ended questions about value capture.
  • Examples:
    • Software developers adopting LLM coding assistants rapidly compared to slower daily consumer habits.
  • Terminology:
    • agentic coding
    • product-market fit
    • LLMs
    • inference
  • Why it matters: Understanding where AI actually creates immediate economic value versus where it remains a solution looking for a problem.

Foundation Models as Commodities and Infrastructure Economics

  • Explanation: Examining whether foundation model providers will maintain pricing power or if models will become commoditized infrastructure.
  • Key claims:
    • Models are increasingly looking like commodities because multiple companies build frontier models doing similar things with similar chips.
    • Unlike operating systems or browsers that captured value through network effects and leverage, foundation models sit low in the stack.
    • A massive capex wave of hundreds of billions to trillions of dollars is flowing into AI infrastructure without clear long-term pricing discipline.
  • Examples:
    • Comparing current AI capex and demand to historical telecommunications and mobile data investments where infrastructure spending outpaced software value capture for years.
  • Terminology:
    • foundation models
    • commodity
    • capex
    • pricing power
    • infrastructure layers
  • Why it matters: Determining who captures the economic surplus of the AI boom—the model builders or the application layer.

Historical Analogies: Mobile, Internet, and PCs

  • Explanation: Using past platform shifts to forecast the trajectory and challenges of the generative AI boom.
  • Key claims:
    • History teaches us that at the early stage of a platform shift, nobody knows how it will work or who the winners will be.
    • Mobile data, cellular networks, and early internet adoption all experienced severe supply/demand crunches, cost imbalances, and over-investment.
    • Just because a technology is revolutionary does not mean early infrastructure builders capture the downstream value.
  • Examples:
    • IBM ad promising '150 extra engineers' using electronic calculators.
    • AT&T launching flat-rate data before streaming and YouTube crushed cellular networks.
    • Telecom operators spending hundreds of billions on capex while cool consumer apps were built by others.
  • Terminology:
    • platform shift
    • cellular networks
    • flat-rate data
    • telecom capex
  • Why it matters: Prevents misplaced certainty by grounding current AI hype in decades of technology adoption cycles.

Enterprise Adoption, Services, and Exception Handling

  • Explanation: Analyzing how enterprises, consultancies, and knowledge-work industries will integrate AI.
  • Key claims:
    • Most enterprise software usage revolves around exception handling and unspoken institutional workflows that aren't documented in training data.
    • Consultancies like McKinsey and BCG sell bespoke, one-off point solutions rather than standardized software products.
    • The fundamental tension in enterprise AI is balancing probabilistic models that make mistakes with deterministic enterprise systems.
  • Examples:
    • Commodities companies using LLMs for cash flow forecasting.
    • Law firms and consultancies deploying custom AI workflows for clients.
  • Terminology:
    • exception handling
    • deterministic systems
    • probabilistic software
    • point solutions
  • Why it matters: Explains why broad enterprise software transformation takes years of organizational friction rather than instant plug-and-play adoption.

Key Points

Foundation models are at high risk of commoditization

  • Explanation: Because multiple labs build comparable frontier models on similar hardware using similar techniques, model providers struggle to maintain lasting pricing power.
  • Evidence: Rapid iteration cycles among OpenAI, Anthropic, open-source models, and hyperscalers.
  • Practical implication: Defensibility and value capture will shift away from raw model training to applications, proprietary data, workflows, and user interfaces.

Historical platform shifts always start with extreme scarcity and confusion

  • Explanation: Every major tech wave—PC, internet, mobile—begins with high costs, capacity crunches, and an inability to predict winning business models.
  • Evidence: Early mobile data pricing crunches and telecom capex cycles in the 2000s.
  • Practical implication: Investors and operators must anticipate painful economic rebalancing between infrastructure costs and end-user willingness to pay.

Enterprise software adoption is constrained by tacit organizational knowledge

  • Explanation: Much of enterprise value creation relies on undocumented workflows and human judgment that probabilistic LLMs cannot instantly automate.
  • Evidence: The parallel between historical consulting advisory models and current AI integration hurdles.
  • Practical implication: True AI transformation requires redesigning organizational workflows, not just plugging an API into existing spreadsheets.

Frameworks, Models & Processes

The Technology Adoption and Value Stack

  • How it works: Analyzes where value accumulates during a technology shift: infrastructure layers versus application and operating system layers.
  • Components:
    • Hardware and Infrastructure capex
    • Foundation Models
    • Developer Tooling and SaaS
    • End-User Applications and Workflows
  • When to use: Evaluating technology investments to see whether infrastructure providers or application layer companies capture long-term profit margins.

Examples & Case Studies

IBM Electronic Calculator ad offering '150 extra engineers'.

  • Illustrates: How every major computing wave frames its early pitch around automating human labor using historical paradigms.
  • Lesson: Early marketing of revolutionary tech often uses old conceptual frameworks to describe new capabilities.

Mobile data traffic surging 1500-2000x while telecom capex stayed flat for 20 years.

  • Illustrates: The divergence between heavy infrastructure investment and downstream value capture.
  • Lesson: Building the underlying pipe doesn't guarantee you capture the economic value created on top of it.

Actionable Takeaways

  • Immediate:
    • Focus AI deployment on high-certainty productivity tasks like software coding assistants where product-market fit is proven.
    • Recognize that foundational model cost curves and capabilities are shifting faster than enterprise procurement cycles can adapt.
  • Strategic:
    • Do not assume foundation model providers will retain pricing power; prepare for model commoditization.
    • Evaluate enterprise AI strategy around workflow redesign rather than simple wrapper applications.
  • Questions to investigate:
    • Where will enterprise profit margins settle as AI inference and training costs drop by orders of magnitude?
    • How will incumbent enterprise software giants absorb or neutralize foundational AI capabilities?
    • At what point does AI capital expenditure outpace sustainable return on investment across hyperscalers?

Claims Worth Verifying

  • Telecom operators spent hundreds of billions annually on capex while revenue remained relatively flat over a 20-year period. (historical economic claim)
  • Software coding is currently the only domain with absolute, validated product-market fit at scale for generative AI. (market analysis claim)

Notable Quotes

"History teaches us nothing except that something will happen." (at 25:41) "AI is amazing and transformative and completely unlike anything that's happened before. Mobile was quite a big deal too." (at 97:35)

Compressed Summary

  • Foundation models risk rapid commoditization due to fierce competition and similar hardware stacks.
  • AI infrastructure capex mirrors historical telecom and mobile data over-investment cycles.
  • Software coding is the primary domain with undeniable product-market fit today.
  • Enterprise adoption is bottlenecked by tacit organizational workflows and probabilistic error handling.
  • Keywords: foundation models, commoditization, capex, infrastructure, product-market fit
  • Core insight: Generative AI foundation models are following the historical path of past platform shifts, transitioning rapidly from scarce technological marvels to commoditized infrastructure.

Core insights

6
Mental Modelmedium noveltymoderate evidence

Because multiple labs build comparable frontier models on similar hardware with similar techniques, foundation models behave like a low-stack commodity without durable pricing power, so defensibility migrates upward to applications, proprietary data, workflows, and UI.

Why it matters

If models are interchangeable inputs, hard architectural investment in model-specific integrations is wasted; engineering leverage moves to the layer that owns data, workflow, and interface.

Generalization

For any commoditizing capability layer, abstract it behind a boundary and compete on the differentiated layer above rather than optimizing for the commoditized layer.

Models are increasingly looking like commodities because multiple companies build frontier models doing similar things with similar chips.
Open source video
Unlike operating systems or browsers that captured value through network effects and leverage, foundation models sit low in the stack.
Open source video
Defensibility and value capture will shift away from raw model training to applications, proprietary data, workflows, and user interfaces.
Open source video
Mechanismhigh noveltymoderate evidence

Most enterprise software value lives in exception handling and unspoken institutional workflows that are not documented in training data, so the automatable slice is smaller than it appears and the remaining exceptions are the hard core, not an edge case.

Why it matters

Agent systems that optimize the happy path will stall at the margin where the real institutional logic lives; exception routing, escalation, and workflow capture are first-class design problems, not error handling.

Generalization

Any automation of a knowledge-work process must explicitly model the undocumented decision exceptions, which are typically the highest-value and least-documented part of the process.

Most enterprise software usage revolves around exception handling and unspoken institutional workflows that aren't documented in training data.
Open source video
True AI transformation requires redesigning organizational workflows, not just plugging an API into existing spreadsheets.
Open source video
Tradeoffmedium noveltymoderate evidence

The core tension in enterprise AI is coupling probabilistic models that make mistakes to deterministic enterprise systems that cannot tolerate them.

Why it matters

Architecture must decide where probabilistic output is contained: deterministic validation, human review boundaries, or downstream compensating controls. This determines reliability and blast radius of failures.

Generalization

Whenever a stochastic component is embedded in a deterministic system of record, the boundary between them is the key reliability-design decision.

The fundamental tension in enterprise AI is balancing probabilistic models that make mistakes with deterministic enterprise systems.
Open source video
Empirical Resultmedium noveltymoderate evidence

Agentic coding is the one place AI has demonstrable product-market fit, evidenced by customers pulling the tool out of the vendor's hands, while most enterprise and consumer use cases remain exploratory.

Why it matters

The clearest validated signal for where to invest engineering effort today is coding assistants; other domains are still searching for value capture and should be treated as experiments, not roadmap certainties.

Generalization

Use demand pull (users actively grabbing the tool) rather than technical capability as the criterion for where to productionize agent systems.

Agentic coding started working immediately, narrowing tech focus onto that specific use case.
Open source video
Software development is the prime example of strong product-market fit because customers actively pull the tool out of your hands.
Open source video
Most other enterprise and consumer use cases remain exploratory with open-ended questions about value capture.
Open source video
Predictionmedium noveltymoderate evidence

A capex wave of hundreds of billions to trillions is flowing into AI infrastructure without clear long-term pricing discipline, and history shows early infrastructure builders frequently fail to capture the downstream value they enable.

Why it matters

Cost of inference and access may be structurally unstable; systems that assume current price/performance trend lines persist are exposed to supply-demand rebalancing.

Generalization

At the start of a platform shift, infrastructure spending and value capture diverge; expect painful economic rebalancing between infrastructure cost and end-user willingness to pay.

A massive capex wave of hundreds of billions to trillions of dollars is flowing into AI infrastructure without clear long-term pricing discipline.
Open source video
Just because a technology is revolutionary does not mean early infrastructure builders capture the downstream value.
Open source video
Predictionmedium noveltyweak evidence

Enterprise AI is currently delivered as bespoke, one-off point solutions by consultancies rather than standardized software products, mirroring historical advisory models.

Why it matters

If enterprise AI is services-shaped today, then custom integration cost, per-customer workflow discovery, and repeatability are the dominant constraints on scaling deployments.

Generalization

When a new capability is delivered as bespoke services rather than productized software, that is a signal the underlying abstraction has not yet been found.

Consultancies like McKinsey and BCG sell bespoke, one-off point solutions rather than standardized software products.
Open source video
Law firms and consultancies deploying custom AI workflows for clients.
Open source video

Deep dives

4

Value capture in the AI stack under model commoditization

Research question

Which layer (hardware, models, tooling, applications) captures durable economic surplus as foundation models commoditize, and how do gross margins compare across layers over successive release cycles?

Why

Determines where engineering and investment should concentrate; if models are interchangeable, model-specific integration is wasted effort and defensibility must be built in applications, data, and workflows.

Models are increasingly looking like commodities because multiple companies build frontier models doing similar things with similar chips.
Open source video
Unlike operating systems or browsers that captured value through network effects and leverage, foundation models sit low in the stack.
Open source video
Defensibility and value capture will shift away from raw model training to applications, proprietary data, workflows, and user interfaces.
Open source video
Source video

Capturing undocumented exception handling and tacit workflows for agent systems

Research question

How can undocumented institutional exception rules be elicited, encoded, and integrated into agent systems to improve real-world enterprise workflow coverage?

Why

Most enterprise value sits in exceptions not present in training data; agents that optimize the happy path will fail at the margin where institutional logic lives, making exception routing a first-class design problem.

Most enterprise software usage revolves around exception handling and unspoken institutional workflows that aren't documented in training data.
Open source video
True AI transformation requires redesigning organizational workflows, not just plugging an API into existing spreadsheets.
Open source video
Source video

Architectural containment of probabilistic model errors in deterministic enterprise systems

Research question

What validation, abstention, and escalation patterns most reliably contain probabilistic model errors before they reach systems of record, and how do they trade off throughput against reliability?

Why

The boundary between stochastic generation and deterministic systems of record is the key reliability-design decision; without a containment layer, model mistakes propagate directly into auditable business processes.

The fundamental tension in enterprise AI is balancing probabilistic models that make mistakes with deterministic enterprise systems.
Open source video
Source video

Demand-pull signals for identifying next product-market fit beyond coding

Research question

Which non-coding use cases exhibit organic demand-pull (users actively pulling the tool) versus capability demos, and what metrics distinguish real product-market fit from exploratory adoption?

Why

Coding is the only proven demand-pull case; treating other domains as roadmap certainties misallocates engineering effort when they may still be searching for value capture.

Agentic coding started working immediately, narrowing tech focus onto that specific use case.
Open source video
Software development is the prime example of strong product-market fit because customers actively pull the tool out of your hands.
Open source video
Most other enterprise and consumer use cases remain exploratory with open-ended questions about value capture.
Open source video
Source video

Article ideas

4

The Commodity Model Trap: Why Foundation Models Won't Capture the AI Surplus

Foundation models are commoditizing because multiple labs build similar capabilities on similar hardware, so defensibility and profit migrate to applications, proprietary data, and workflows; companies betting on model lock-in are misallocating engineering effort.

Angle

Argues against model-centric strategies and for architectural abstraction of model access, using historical platform shifts as evidence.

Source video

The Exception Is the Product: Why Enterprise AI Fails on the Happy Path

Most enterprise value lies in undocumented exception handling and tacit workflows, so agent systems that optimize standard paths will fail; the winning architecture treats exception routing and workflow capture as core features, not error handling.

Angle

Reframes exception handling from an edge case to the central design problem, with concrete architectural implications for agent systems.

Source video

Probabilistic Output, Deterministic Consequences: Designing the Boundary Layer

The key reliability design decision in enterprise AI is where to contain probabilistic outputs; without a deterministic validation and escalation layer, model errors propagate into systems of record, and reliability requirements should drive architecture at the model/system boundary.

Angle

Provides a design framework for validation, abstention, and escalation between LLMs and systems of record.

Source video

Demand Pull, Not Demos: Finding the Next Agentic Coding

Coding assistants succeeded because users pulled the tool; most other AI use cases are capability demos, so we need demand-pull metrics (organic adoption, unsolicited requests) to distinguish real product-market fit from exploratory investments.

Angle

Proposes a demand-pull evaluation framework and warns against over-concentrating on one proven use case while others remain unvalidated.

Source video

Project ideas

4

Model Abstraction Router

new

Abstracting model access behind a routing layer that selects models per task based on cost/performance reduces vendor lock-in and improves cost efficiency without degrading task success compared to a single-frontier-model baseline.

Proof of concept

Build a router that sends requests to different frontier models (e.g., GPT-4, Claude, Gemini) based on task type and measures cost/quality versus a single-model baseline across a benchmark suite of coding, reasoning, and enterprise tasks.

Measurement

Cost per successful task, latency, and accuracy/quality scores across benchmark tasks.

Source video

Exception Miner

new

Explicitly eliciting exception rules from operators and encoding them as structured context improves agent coverage on enterprise workflows more than fine-tuning on standard-path data alone.

Proof of concept

Build a pipeline that interviews operators, extracts exception rules (e.g., 'if invoice > $10k, require manager approval'), and injects them into an agent for a sample workflow like invoice processing or IT ticket triage.

Measurement

Reduction in escalated cases, increase in straight-through processing rate, and accuracy on exception cases versus a baseline without exception rules.

Source video

Probabilistic Gate

gatehouse

A deterministic validation and escalation layer between LLM output and systems of record reduces catastrophic errors without unacceptable throughput loss.

Proof of concept

Implement a validation layer with rule-based checks, confidence thresholds, and human-in-the-loop escalation for a sample task such as generating SQL queries or API calls that write to a database.

Measurement

Error rate reaching system of record, false positive/negative rates, latency overhead per request, and human escalation rate.

Source video

Demand Pull Tracker

new

Tracking organic demand-pull signals (unsolicited adoption, user-initiated requests, organic retention) across candidate AI use cases predicts which domains will achieve product-market fit better than capability benchmarks.

Proof of concept

Instrument a set of enterprise AI pilots to log demand-pull metrics (e.g., frequency of unprompted usage, feature requests, expansion requests) and correlate with 6-month retention and expansion.

Measurement

Correlation between early demand-pull signals and later retention/expansion, and predictive power versus benchmark scores.

Source video

Architectural implications

3

Foundation models behave as a commodity low in the value stack.

Before

Architectures assume lock-in to a single frontier model provider and optimize prompts/features against that specific model.

After

Model access is abstracted behind an interchangeable boundary, with routing and selection treated as an operational concern rather than an application-level commitment.

Consequence

Engineering effort concentrates on proprietary data, workflow, and interface layers that actually hold defensibility.

Source video

Enterprise value sits in undocumented exception handling and tacit workflows.

Before

AI integration is designed as plugging an API into existing tools and spreadsheets.

After

The system explicitly models exception paths, escalation to humans, and captures the unspoken institutional rules as structured context.

Consequence

Adoption becomes a workflow-redesign effort with human-in-the-loop boundaries, increasing integration cost but improving real coverage.

Source video

Probabilistic model outputs feed deterministic enterprise systems.

Before

Model output flows directly into systems of record.

After

A deterministic validation and control layer mediates between probabilistic generation and deterministic downstream actions.

Consequence

Reliability and failure containment become design requirements at the model/system boundary.

Source video

Tradeoffs and failure modes

4

Infrastructure capex vs pricing discipline

Benefit

Massive capex builds capacity and drives rapid capability improvement across many competing model builders.

Cost or risk

Spending without clear long-term pricing discipline creates supply/demand imbalances and cost burdens that early infrastructure builders may not recoup.

A massive capex wave of hundreds of billions to trillions of dollars is flowing into AI infrastructure without clear long-term pricing discipline.
Open source video
Source video

Bespoke point solutions vs standardized product

Benefit

Consultancy-style, one-off point solutions fit each client's undocumented workflows and can be deployed immediately.

Cost or risk

They do not scale as reusable software, keeping enterprise AI in a services mold with high per-customer cost.

Consultancies like McKinsey and BCG sell bespoke, one-off point solutions rather than standardized software products.
Open source video
Source video

Probabilistic models in deterministic enterprises

Benefit

LLMs can address fuzzy, judgment-heavy tasks that rule-based systems cannot.

Cost or risk

Model mistakes collide with deterministic systems that require exact, auditable behavior.

The fundamental tension in enterprise AI is balancing probabilistic models that make mistakes with deterministic enterprise systems.
Open source video
Source video

Narrow focus on coding as the proven use case

Benefit

Focusing on coding yields the strongest demand pull and clearest product-market fit.

Cost or risk

Over-concentration on one use case narrows tech focus and leaves broader enterprise and consumer value capture still unresolved.

Agentic coding started working immediately, narrowing tech focus onto that specific use case.
Open source video
Source video

Open questions

4

Which layer of the value stack captures the durable economic surplus from the AI boom?

Why unresolved

The summary states that foundation models sit low in the stack and that revolutionary technology does not imply infrastructure builders capture downstream value, but does not identify the eventual winner.

Research direction

Model value-stack economics per layer (hardware, models, tooling, applications) and compare margin capture patterns against prior platform shifts.

Source video

How can undocumented institutional workflows and exception logic be captured into agent systems when they are absent from training data?

Why unresolved

The summary identifies exception handling and tacit knowledge as the core constraint but offers no mechanism for extracting or encoding it.

Research direction

Prototype workflow-capture pipelines that elicit exception rules from operators and encode them as structured context or deterministic guardrails.

Source video

What architectural boundary reliably contains probabilistic model errors before they reach deterministic enterprise systems?

Why unresolved

The tension is named but no containment pattern, validation method, or escalation design is specified.

Research direction

Design and benchmark validation/abstention/escalation layers at the model-to-system-of-record boundary.

Source video

Which non-coding use cases will show the 'customers pull the tool out of your hands' demand-pull signal?

Why unresolved

The summary says most other enterprise and consumer use cases remain exploratory with open questions about value capture.

Research direction

Track demand-pull indicators (organic pull, unsolicited adoption) across candidate domains to distinguish real PMF from capability demos.

Source video

Key claims

8
causalVerification needed

Multiple companies build comparable frontier models on similar hardware and techniques, so models are becoming commodities.

Evidence

Models are increasingly looking like commodities because multiple companies build frontier models doing similar things with similar chips.

Question

Do measurable capability, cost, and switching-cost gaps between frontier models persist or compress over successive release cycles?

Source video
comparativeVerification needed

Unlike operating systems or browsers, foundation models cannot capture value through network effects and leverage because they sit low in the stack.

Evidence

Unlike operating systems or browsers that captured value through network effects and leverage, foundation models sit low in the stack.

Question

Do foundation model providers exhibit any demand-side economies of scale comparable to OS or browser network effects?

Source video
predictionVerification needed

Value capture will shift away from raw model training toward applications, proprietary data, workflows, and user interfaces.

Evidence

Defensibility and value capture will shift away from raw model training to applications, proprietary data, workflows, and user interfaces.

Question

Do gross margins of application-layer AI companies exceed those of model-training-layer companies over time?

Source video
causalVerification needed

Software development is the prime example of AI product-market fit because customers actively pull the tool out of your hands.

Evidence

Software development is the prime example of strong product-market fit because customers actively pull the tool out of your hands.

Question

What measurable adoption and retention metrics distinguish coding assistants from other enterprise AI deployments?

Source video
factualVerification needed

Most enterprise software usage revolves around exception handling and unspoken institutional workflows not documented in training data.

Evidence

Most enterprise software usage revolves around exception handling and unspoken institutional workflows that aren't documented in training data.

Question

What fraction of enterprise workflow steps are exception paths versus standard-path steps in representative deployments?

Source video
causalVerification not requested

There is a fundamental tension between probabilistic models that make mistakes and deterministic enterprise systems.

Evidence

The fundamental tension in enterprise AI is balancing probabilistic models that make mistakes with deterministic enterprise systems.

Source video
factualVerification needed

A capex wave of hundreds of billions to trillions is flowing into AI infrastructure without clear long-term pricing discipline.

Evidence

A massive capex wave of hundreds of billions to trillions of dollars is flowing into AI infrastructure without clear long-term pricing discipline.

Question

What are aggregate AI infrastructure capex figures and how do they compare to current and projected inference/API revenue?

Source video
causalVerification needed

Just because a technology is revolutionary does not mean early infrastructure builders capture the downstream value.

Evidence

Just because a technology is revolutionary does not mean early infrastructure builders capture the downstream value.

Question

In past platform shifts (telecom, mobile data, early internet), what share of downstream value was captured by infrastructure builders versus application builders?

Source video

Connections

5